Low Illumination Object Detection using Deep Learning Techniques.

Abstract

Since object detection in the actual world is affected by illumination, object detection has been a challenging task for quite a long time. Under normal lighting conditions, image processing achieves significant efficiency much of the time. In low-light situations, however, a picture becomes blurry and dim, making subsequent computer vision tasks difficult and it has a significant impact on our vision’s performance. While research on low-light images has grown steadily, especially in the area of low light, there is still a need for further breakthroughs. Recent research suggests that low light is more difficult for machine cognition than initially expected. Challenges of object detection in low light were discussed in our research. To begin, a high-quality largescale ExDark dataset of dynamic scenes shot at night is used. Following that, the roboflow software was used to annotate the dataset it was used to give labels to every image. A baseline model was offered for performance study to identify areas for future research and demonstrate that low light is a nontrivial challenge that necessitates special attention from researchers. Object detector was used to improve the accuracy of low light objects. You Only Look Once (YOLO) is an as of late proposed bound together object identification approach that can straightforwardly regress from input picture to predict classes and loacation. A general applicability of YOLOV4 on the ExDark Dataset is tested in this study. Thoroughly examined detectors performance from all angles and accuray was improved. MaP was used as an accuracy measurement for object detectors. The proposed method enhances detection accuracy by 74 percent based on the results of testing. The predicted findings can be used as a beginning stage for additional research on low-light object detection. Finally, current state of research is described, with recommendations for future study areas.